spent 6 weeks building a prototype and still can't figure out my data pipeline

So I've been heads-down on a small AI tool for about six weeks now. Built the prototype in Cursor — my coworker suggested it because she said the AI completion quality was noticeably better for long contexts, and honestly she wasn't wrong. It's doing something with contract analysis, which sounds boring but the pattern-matching piece is actually working better than I expected.

Here's where I'm stuck. My current pipeline is duct-taped together: I'm pulling from three different data sources, doing some preprocessing in Python, then sending chunks to the API. Around 37 users have signed up from a cold post I made on a forum, and they seem to genuinely want this. That part is encouraging.

The preprocessing step is killing me though. Right now I'm manually reviewing outputs before they go to users, which means I can't scale without it becoming a second job. I've looked at a few approaches but none feel right. I could automate more of the review step but I keep second-guessing whether I'm just avoiding harder architecture work.

I'm vaguely considering Supabase for the storage layer — saw someone mention it handles JSON payloads well — but I haven't touched it yet and the migration sounds painful.

For those of you who went from prototype to something you'd actually call a product, how did you handle the messy middle stage where manual processes are holding everything together? Is there a point where you just accept the duct-tape for a while versus rip it out?